Our manufacturing plant is looking to automate our QA process for detecting microscopic cracks in metal components. We are debating between traditional sensor-based systems and a Computer Vision (CV) model using Deep Learning. Is CV reliable enough to handle varying lighting conditions and the high speed of an assembly line, or are we better off sticking with ultrasonic testing for now? What kind of hardware (GPUs/Edge devices) would we realistically need?
3 answers
Computer Vision has evolved significantly for industrial use. In a project I worked on in early 2024, we used "Synthetic Data" to train the model on millions of simulated lighting conditions, which made it incredibly robust. For an assembly line, you don't want to send data to the cloud; you need "Edge AI." We deployed our models on NVIDIA Jetson devices right on the factory floor. The latency was near zero, and the accuracy for surface cracks actually surpassed human inspectors. However, for internal cracks, you’d still need ultrasonic—CV only sees what the camera can capture.
What was the "False Discovery Rate" like when you first started, and how long did it take to calibrate the model for the floor environment?
Look into "Anomaly Detection" models like Autoencoders. They are great because they only need to see "perfect" parts to identify anything that looks "weird."
Great point, Thomas! Using Autoencoders for anomaly detection is much faster than traditional object detection when you have a million different ways a part could be "broken."
Richard, the initial FDR was around 8%, which was too high for production. It took us about 6 weeks of "Active Learning"—where the human inspector corrected the AI's mistakes—to bring it down to 0.5%. The trick is to have a feedback loop where the model is constantly retrained on the "hard cases" it fails on. It’s not a "set it and forget it" technology, but once it’s calibrated, the ROI in saved time and materials is massive.